Decentralized Dynamic ADMM with Quantized and Censored Communications
Decentralized Dynamic ADMM with Quantized and Censored Communications
复制标题
DOI：
10.1109/ieeeconf44664.2019.9048719
复制
发表时间：
2019-11
期刊：
影响因子：
--
通讯作者：
Yaohua Liu;K. Yuan;Gang Wu;Z. Tian;Qing Ling
中科院分区：
文献类型：
--
作者：
Yaohua Liu;K. Yuan;Gang Wu;Z. Tian;Qing Ling
本文提出了一种量化和通信删失交替方向乘子方法(ADMM)来求解一个定义在分散网络上的动态优化问题。在每一时刻，网络节点协作最小化其时变的局部目标函数的总和。该算法称为DQC-ADMM，通过局部迭代计算和通信来跟踪时变的最优解。不同于传统方法需要节点每次都向邻居发送精确的局部迭代，我们对发送的局部迭代进行量化，并引入通信审查机制，以减少优化过程中的通信开销。虽然量化和删失通信造成了不准确，但我们严格证明了所提出的算法能够在温和的条件下以有界误差跟踪时变最优解。并通过数值实验验证了该算法的跟踪性能和通信节省。
In this paper, we develop a quantized and communication-censored alternating direction method of multipliers (ADMM) to solve a dynamic optimization problem defined over a decentralized network. At every time, the network nodes collaboratively minimize the summation of their local objective functions, which are time-varying. The proposed algorithm, termed as DQC-ADMM, tracks the time-varying optimal solution through local iterative computation and communication. Unlike traditional approaches that require the nodes to transmit the exact local iterates to their neighbors at every time, we quantize the transmitted local iterates and introduce a communication-censoring mechanism so as to reduce the communication cost spent in the optimization process. Although the quantized and censored communications cause inaccuracy, we rigorously prove that the proposed algorithm is able to track the time-varying optimal solution with a bounded error under mild conditions. We also demonstrate the tracking performance and communication savings of the proposed algorithm through numerical experiments.